AgentVoice is a bridge from your voice portal to your desired coding agent (claude code, cursor, codex) via an custom, global mcp server and more granular cli commands specific to the provider, such as new agent. For wake word detection, it uses a browser loaded, very lightweight vosk model, or a VAD model (silero VAD) instead of outsourcing the entire stream to speech to text provider. It does not actually use speech to speech models, this is a design choice, I intentionally wanted the coding agent provider to handle both user communication and actual implementation for quality responses and context, it works decently well in practice. This bridge can be run locally for testing, but also hosted, for remote accessibility from the device itself, I used tailscale for this, but other methods, such as headscale are also supported and encourage, as to reduce reliance on closed tailscale servers and have more control and privacy over your data. Interestingly, for text to speech, it was found that nearly every single modern web browser offers this for free, had to be taken advantage of. For speech to text however, I will note that just temporarily only Amazon Bedrock was supported for transcription, as this was a hobby project and it was the only provider accessible to me, but I am working on expanding support to other providers as well, such as OpenAI whisper, or Groq, and even completely self hosted for a professional setup and if you have premium hardware accessible. My dad and I both use this on a regular basis, especially where typing on tiny keyboards and interacting with the screen is not efficient, dangerous, and a regular PC is not available (e.g. driving to work and back). This is just a brief summary, for details check out the project, from experience, worth trying out, please consider contributing as well!
AgentVoice is a bridge from your voice portal to your desired coding agent (claude code, cursor, codex) via an custom, global mcp server and more granular cli commands specific to the provider, such as new agent. For wake word detection, it uses a browser loaded, very lightweight vosk model, or a VAD model (silero VAD) instead of outsourcing the entire stream to speech to text provider. It does not actually use speech to speech models, this is a design choice, I intentionally wanted the coding agent provider to handle both user communication and actual implementation for quality responses and context, it works decently well in practice. This bridge can be run locally for testing, but also hosted, for remote accessibility from the device itself, I used tailscale for this, but other methods, such as headscale are also supported and encourage, as to reduce reliance on closed tailscale servers and have more control and privacy over your data. Interestingly, for text to speech, it was found that nearly every single modern web browser offers this for free, had to be taken advantage of. For speech to text however, I will note that just temporarily only Amazon Bedrock was supported for transcription, as this was a hobby project and it was the only provider accessible to me, but I am working on expanding support to other providers as well, such as OpenAI whisper, or Groq, and even completely self hosted for a professional setup and if you have premium hardware accessible. My dad and I both use this on a regular basis, especially where typing on tiny keyboards and interacting with the screen is not efficient, dangerous, and a regular PC is not available (e.g. driving to work and back). This is just a brief summary, for details check out the project, from experience, worth trying out, please consider contributing as well!